Meta’s Muse Code Fans Out Six Agents at Once

Meta has launched Muse Code, a terminal-based coding agent aimed squarely at the workflow Anthropic's Claude Code and OpenAI's Codex have spent the past year defining. It is in beta, it installs with a single command, and it is powered by Meta's own Muse Spark coding model. TechCrunch reported the launch on August 5.

The pitch is scale rather than novelty. Muse Code is built for large repositories, and its headline mechanic is fan-out: instead of working a task front to back in one context, it spawns sub-agents that run at the same time, each in its own isolated worktree. Mark Zuckerberg described the agent as able to "fan out to separate sub-agents working in parallel in isolated worktrees." In Meta's own demo, that meant building six game features concurrently without the agents overwriting each other's files.

Isolation is the part that matters technically. Parallel agents editing one checkout is a merge-conflict generator; giving each a separate worktree turns six racing writers into six independent branches that can be validated and reconciled afterward, and leaves the developer's original tree untouched. Meta says the agent handles the full loop on each of them — planning the change, writing the code, then validating the result.

On positioning, Meta is not claiming the frontier. Alexandr Wang, the company's AI chief, framed Muse Code as "an incredibly good option, especially from a cost perspective" across many workflows — a price argument, not a capability one. That is a coherent read of where agentic coding has landed: the tools are close enough on routine work that per-token cost decides which one a large engineering org standardizes on.

There is an awkward piece of recent history attached. In June, Meta barred its own engineers from using Claude Code and Codex, a restriction framed at the time around rival providers' terms on model distillation. With Muse Code shipping seven weeks later, the same policy also has the effect of clearing the field for an in-house tool whose first serious users are Meta's own engineers.

Meta has not published benchmark numbers, pricing, or general-availability timing, and beta terminal agents are exactly the category where demo fluency and repository reality diverge. The six-features-at-once demo is a good showcase of the orchestration layer; it says less about whether Muse Spark holds up on an unfamiliar million-line codebase with bad tests and twelve years of undocumented decisions. That is the benchmark that will actually decide this one, and nobody has run it yet.